Lightweight Neural Network Ensemble Models for Medical Image Classification with MedMNIST Dataset

Over the last decade, convolutional neural networks (CNNs) have demonstrated significant potential in assisting medical diagnosis. Research has focused on implementing such networks, particularly in the context of embedded medical systems. This can be challenging due to CNNs’ large memory footprint, high computational requirements, and need for retained performance. One of the common approaches for high model accuracy is to use an ensemble of several deep neural networks (DNNs). Recently, the authors have discussed the high computational demands of DNN ensembles. Therefore, in this paper, lightweight ensemble solutions are investigated. Three separate types of ensembles are classified: DNN-only (consisting of deep-of-the-shelf networks), CNN-only (consisting of customized CNNs), and a hybrid ensemble (combining the former two architectures). Experiments were conducted on each class using three public datasets from the MedMNIST database, and classes were compared and contrasted. The results show that a higher sensitivity and a smaller memory footprint can be achieved with CNN-only compared to with DNN-only. Moreover, a hybrid ensemble approach is proposed as the best compromise between the two, being the most lightweight, as it reduces the number of FLOPs, with the performance result comparable to previous work. The performance drop is ∼0.3%, ∼0.4%, and ∼2% for PathMNIST, OrganAMNIST, and OCTMNIST, but the memory footprint is reduced by ∼65%, ∼77%, and ∼82% compared to the recent state of the art. Thus, the proposed hybrid ensemble approach is compliant with the requirements of a resource-constrained device and suitable for implementation in a smart medical system.